Google's AI chiefs leave to build agents, Karpathy joins Anthropic

Google DeepMind CEO Demis Hassabis becomes Chair and Alphabet Chief Scientist as Jeff Dean, Sanjay Ghemawat, and two colleagues leave Google to found Discovery Loop: On the same day, Google announced two seismic leadership changes. Alphabet CEO Sundar Pichai confirmed that Demis Hassabis is handing over day-to-day operational responsibilities at Google DeepMind to Koray Kavukcuoglu, who becomes SVP. Hours later, Jeff Dean and Sanjay Ghemawat, two of the engineers most identified with Google's AI and infrastructure dominance over the past 27 years, announced they are leaving alongside researchers Oriol Vinyals and Quoc Le to co-found Discovery Loop, a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering research by parallelising thousands of experimental loops using frontier AI models.

Meta launches Muse Code in beta, a terminal coding agent powered by Muse Spark 1.2: Muse Code handles full software-engineering tasks across large repos by running persistent background agents that accumulate context across a session, fanning out to parallel sub-agents in isolated worktrees for bigger jobs so the working copy is never touched. It is Meta's second model release in under four weeks and benchmarks show Muse Spark 1.2 is competitive with top coding models, though some observers note the published evals omit several leading competitors.

Prime Intellect launches Prime Agent, a self-improving RLM harness that hits 95.5% on ARC-AGI-3: Prime Agent wraps frontier models in a "Recursive Language Model" harness, giving the model a persistent Python REPL as its only tool: it can rewrite its own context, spawn sub-agents, and modify harness state at runtime. Applied to Claude Opus 4.5, the same model that scored 30.2% with its native harness jumped to 95.5% on ARC-AGI-3, above the human-expert baseline. One reply notes the comparison may flatter Prime Agent because Codex with a minimal prompt also reaches ~78%, not the 13.3% headline figure, so the gain versus proprietary harnesses is real but contested. The open-source repo is MIT-licensed and works across models.

Google in talks to acquire Mechanize for $1.5B, its biggest AI data deal since Scale (paywalled): Mechanize, founded in April 2025 by ex-Epoch AI researchers, pays a team of 35-50 engineers roughly $400K/yr each to build high-quality reinforcement learning coding environments, producing about one task per week at ~$8K per task for use by frontier AI labs. The startup raised $9.1M at a $500M valuation in April 2026, making a potential $1.5B deal a 3x jump in three months. The reported structure is a talent-and-license arrangement, similar to Google's Windsurf and Character.​AI deals, and acquiring the pipeline also denies it to competitors.

Ilya Sutskever's Safe Superintelligence is reportedly preparing its first model release this month: The claim comes from Atreides Management CIO Gavin Baker, speaking on the Invest Like the Best "AI Market Jitters" episode, not from SSI itself. SSI's own website declares its sole focus is building "one goal and one product: a safe superintelligence," which makes even a preliminary commercial release noteworthy. SSI has not confirmed a date, so treat this as an unverified investor-sourced signal.

OpenAI researcher Jeffrey Wang explains why reinforcement learning has become an inference workload: In a Big Chip Club interview, OpenAI researcher Jeffrey Wang argues that RL training now requires co-designing training and inference pipelines together, because generating rollouts at scale turns RL into a heavy inference workload. It's a concise first-person explanation of the architectural shift driving demand for faster inference hardware.

Anthropic's eval-agent methodology, Stanford's AutoMem 32B paper, and Andrew Ng on loops and graphs, the self-improving agent playbook: Three threads converge on the same idea: evals and memory, not bigger models, unlock self-improving agents. Demystifying evals for AI agents, Anthropic's engineering blog post, formalises the approach, start with 50 real prompts, target 50% pass rate (too easy means the eval is useless), and treat every failed transcript as a training signal. The Stanford AutoMem paper shows a 32B open model reaching frontier-level performance on long-horizon tasks by teaching the agent to manage its own memory (read, write, search, forget), cutting redundant writes by up to 83% and per-step memory by 95%, without touching model weights. Stanford's CS329A course on self-improving AI agents, co-taught by instructors who built Claude and Gemini, covers the research behind these patterns. Andrew Ng's talk on the future of agentic AI makes the case for loops and graphs as the next structural primitive.

Moonshot AI's Kimi K3 arrives as the world's first open 3T-class model, with a 1M-token context window: The 2.8-trillion-parameter mixture-of-experts model from Chinese startup Moonshot AI ships with native vision, tool calling, and reasoning support, making it the largest open-weight release to date. ML engineer Santiago Valdarrama called it the best open-weight model he has tested, noting the size alone explains why closed-source labs are paying attention. Full weights are open, though self-hosting at this scale remains impractical for most teams, so the real pressure is on API pricing.

Devin Outposts lets cloud agents build, sign, notarize, and DMG-package a native SwiftUI macOS app end-to-end in a real Mac environment: Cognition's new Devin Outposts feature routes Devin Cloud sessions onto any machine you operate, including real Mac hardware, so the agent can handle notarytool signing, ticket stapling, and DMG packaging steps that previously required manual intervention. Platform partners including Namespace offer macOS devboxes for teams without physical Apple hardware.

Greptile v5, ground-up rewrite delivers 2x faster code review, 29% more bug catches, and 14pp precision gain: Greptile CEO Daksh Gupta rebuilt the agent architecture around a swarm of parallel agents each narrowly scoped to a single bug hypothesis, cutting end-to-end review time to roughly two minutes. Measured on over a million live production pull requests, v5 raises "great catch" responses from 0.266 to 0.342 per PR and strict address rate from 52% to 66%, with all users upgraded automatically.

Wispr Flow Notetaker is a silent, bot-free meeting recorder that pipes transcripts into AI tools via MCP: The Mac app captures meetings without dropping a bot into the call, attributes each line to the correct speaker, and exposes the full history through MCP so Claude, ChatGPT, Cursor, or any MCP-compatible tool can query past meetings as live context rather than leaving notes to rot in a folder.

Warp Agent CLI launches as a standalone coding agent that works in any terminal: Warp CEO Zach Lloyd introduced the Warp Agent CLI, a coding agent designed to be used outside of the Warp terminal itself, running in any shell environment. The pitch is orchestration-first: the agent can change directories, coordinate multi-agent workflows, and route to cloud execution, capabilities that existing CLI coding agents handle poorly.

Claude Code 2.1.223 adds GitHub org wildcard permissions and fixes invisible-Unicode permission bypass: The release adds "owner/*" wildcard entries to marketplace managed settings so admins can allow or block all repos under a GitHub org in one policy. It also patches a security flaw where commands padded with tabs or invisible Unicode characters could hide portions of themselves from Claude Code's permission approval dialogs, meaning users may have unknowingly approved hidden commands in earlier versions.

LangChain launches LangSmith Engine to automate agent debugging and fixes in production: Getting an agent to production is slower than building it, because engineers must manually read traces, spot failure patterns, and ship patches. LangSmith Engine automates that loop: it continuously clusters production failures into named issues, diagnoses root causes against your codebase, and drafts PRs and evaluators for human review. Each resolved issue also strengthens the eval suite by pulling failing traces into an offline dataset. One caveat worth noting: the same failed trace can support several plausible patches, so reviewers should apply acceptance tests before merging any auto-generated fix. A free quickstart course walks through the full Agent Development Lifecycle.

DSPy Flex lets optimizers rewrite your program's code, not just its prompts: DSPy's new [dspy.​Flex](https://dspy.​ai/diving-deeper/flex/) module is a drop-in replacement for Predict, ReAct, or RLM that hands the GEPA optimizer full control over the Python code itself, not just the instructions. In one benchmark task, pairing Flex with GEPA pushed accuracy from 90% to 95% while cutting LLM calls by roughly 75%, because the rewritten code avoids redundant inference steps entirely.

Andrej Karpathy joins Anthropic to build a team using Claude to accelerate Claude's own pretraining research: OpenAI co-founder and former Tesla AI director Andrej Karpathy has joined Anthropic's pretraining team, where he will lead a group focused on using Claude to speed up the research that produces the next generation of Claude itself. Anthropic's head of pretraining Nicholas Joseph confirmed the hire, and an Anthropic spokesperson told VentureBeat the move pushes toward using AI to help develop AI at the frontier level.

PostHog explains how AI agents are driving its goal of 10,000 pull requests a month: PostHog's engineering blog details a dramatic shift in how software gets built: PRs opened per month jumped from 1,441 in January to 4,725 in June with only ~10% more headcount, and roughly 70% are now opened by AI agents. The company's declared north star is "10,000 PRs a month is easy," potentially reached as soon as September. This accompanies a broader strategic pivot, documented in a separate post, toward making products "self-driving," where agents diagnose problems and generate fixes automatically rather than waiting for engineers to prompt them.

ChainDrop npm worm hits keyv/cacheable ecosystem, poisons 440+ packages and 2,200 versions in under four hours: A self-replicating worm dubbed ChainDrop compromised a maintainer's GitHub account to inject credential-stealing malware starting with [email protected], then spread autonomously to over 440 dependent packages. The payload exfiltrates GitHub tokens, AWS keys, SSH keys, and.​env files from CI/CD runners, disguises itself as a fake "claude" commit with the message "chore: update config", and retrieves command-and-control domains via an Ethereum smart contract. Packages including flat-cache and file-entry-cache are affected. Running npm config set ignore-scripts true --global blocks the install-time execution vector while audits are ongoing.

Turbopuffer launches namespace sharding beta, enabling search indexes up to 256 TB in a single namespace: The object-storage-native search database now lets teams index massive corpora (think web-scale or full-enterprise data lakes) without splitting them across multiple namespaces. Sharding distributes documents across internal shards via hash(id), while the namespace stays a single logical unit for writes and queries.

Supabase open-sources Evals, a benchmark for how well AI coding agents build with Supabase: Supabase is cementing its position as the go-to database for AI-generated apps by releasing an open-source eval suite that runs Claude Code, Codex, and OpenCode against real Supabase tasks (schema design, Edge Functions, RLS policies) and scores them. The move reflects a broader observation that AI coding tools increasingly recommend Supabase by default when users ask what database to use, a dynamic some compare to paying to be the default search engine, except here it is won through training data and developer tooling rather than contracts.

Cloudflare OS: an open platform for agents, apps, and work: Cloudflare has open-sourced Cloudflare OS, a platform that lets teams build AI agents, internal apps, and automated workflows on top of the company's existing global network. Unsupervised Learning founder Daniel Miessler flags it as the latest move in Cloudflare's long-running effort to quietly become the default infrastructure layer for AI, building on years of incremental bets that have steadily expanded its role well beyond CDN and DNS.

Anthropic makes Claude Cowork ready for enterprise as four agentic products converge on the knowledge economy: Anthropic's April 2026 announcement details enterprise-grade role-based access controls, admin analytics, and compliance features for Claude Cowork, its Code-style agent aimed at non-technical knowledge workers. Signull Labs founder @​signulll argues the four products taken together, Claude Code, Claude Cowork, ChatGPT Work, and OpenAI Codex, make most white-collar work either already automatable or visibly next, and that you could walk into any small business today and automate at least 50% of it. Replies are split, with some calling the 50% figure too bullish and others pointing to real trust gaps and compliance barriers, though the directional claim is hard to dismiss. ChatGPT Work, powered by GPT-5.6, ships as a cross-app agent that can stay with complex projects for hours, and no other vendor is close in capability or cultural momentum.

Chamath Palihapitiya argues bottoms-up GTM will be the worst strategic decision of the past decade: Social Capital CEO Chamath Palihapitiya contends that AI will rapidly clone every bottoms-up SaaS tool, making the model obsolete on two fronts: enterprises will treat the resulting tool sprawl as an AI-sovereignty risk (random employees feeding company data into point-solution AIs), and CFOs will simply block the spend with corporate-card filters, pushing buying decisions back to centralized, top-down procurement.

Robert Martin's rule for building with agents: if it can be deterministic, make it deterministic code: Clean Code author Robert C. Martin argues that anything with a predictable, repeatable output should be handled by a deterministic tool or script, not left to an LLM agent. Agents should be reserved for genuinely unstructured tasks where non-determinism is unavoidable. Writing a 500-word system prompt to parse a string or do math, instead of four lines of code, wastes tokens and degrades quality.

Telling a model to "think hard" outperforms setting it to high reasoning mode: Raindrop co-founder and CTO Ben Hylak ran a small but reproducible benchmark showing that adding the words "think hard" to a user message on Claude Opus 5 at low reasoning increased total tokens by 66%, beating the plain high-reasoning baseline. The effect holds on OpenAI's Sol model too, and placing the instruction in the user message works better than the system prompt. The provider-level reasoning "effort knob" (low/medium/high) turns out to be a coarser lever than natural-language instruction.

NeetCode founder Navdeep Singh questions whether frontier-lab "slow down" calls are really about safety: Singh argues the recursive self-improvement risk cited by US AI labs is not even theoretically proven, and that the more straightforward explanation is that training costs have hit a breaking point while Chinese open-source models close the gap fast. The safety framing, he suggests, is doing convenient work for the economics.

Opus 5 backlash grows as users report lazy, unreliable responses in real codebases: A thread started by AI newsletter writer Kim Isenberg, drawing over 1,200 comments with near-unanimous complaints, captures a wave of user frustration: Opus 5 skips unread emails, hallucinates task completion, and reads from memory instead of the actual code. Michael Timbs calls it "a huge liability" for production codebases, and Miles Deutscher summarizes the mood as going "from Fable 5 euphoria to pure despair in a matter of weeks." Some users say Fable 5 is also degraded, frequently reverting to Opus 5 mid-session, and Anthropic has offered no public statement.

DeepSeek-V4-Flash-0731 is Ollama's fastest-growing model ever by token usage, and Vercel AI Gateway cuts its cost 90%+ with one line of code: Ollama is scaling US and Europe capacity to handle demand for the 284B-parameter MoE model, which runs at 100+ tokens/second with no data retention on its cloud. Vercel's AI Gateway changelog shows that routing to Novita via the AI SDK's order option drops the price 90% for Pro customers through August 11.

Higgsfield open-sources the full production package for its $500K AI feature film Hell Grind ahead of a $1M Global Film Festival: Every prompt and asset from the 95-minute action-fantasy film is now public, including 3,000-word-per-shot prompts, lighting notes, and top-down action maps used to keep complex scenes coherent. The production generated 16,181 clips to assemble the first 25 minutes alone (roughly 64 attempts per keeper). Variety's review found the visuals impressively consistent but noted lingering AI tells: flat villain expressions, stiff dialogue, and prop glitches. Hell Grind screened at the Cannes Marché du Film (the paid commercial market, not the main festival competition). Higgsfield is tying the release to its $1M Global Film Festival contest, where creators can now use the full pipeline as a reference.

Harness is so so important, model choice matters less than the scaffolding around it: heybossAI founder Xiaoyin Qu sparked a high-engagement debate arguing that DeepSeek V4 Flash underperforms not because the model is weak but because most developers lack a high-quality agent harness like Codex or Claude Code to wrap it. DAIR.​AI founder Elvis Saravia points to the new DataSpace benchmark, a 410-task, 15 GB evaluation of data agents across heterogeneous file types, as evidence that harness design drives measurable differences in agent outcomes. LangChain Labs applied research lead Vivek Trivedi dissents, calling a "universal harness" a technical myth and arguing that a small set of accepted patterns (compaction, subagents, bash, filesystem) covers most economically valuable tasks, with most harness complexity being strippable.

Bending Spoons acquires Airtable for $1.285B as founder Howie Liu spins out AI-agent startup Hyperagent: Italian software conglomerate Bending Spoons struck an all-cash deal valuing Airtable's operating business at $1.285B enterprise value (roughly $2.25B including net cash), a steep drop from its $11B+ peak. Before signing, Airtable co-founder and CEO Howie Liu carved Hyperagent, his AI-agent platform, into a separate company that sits entirely outside the acquisition. Hyperagent lets users hand open-ended briefs to fleets of agents that research, build, and keep deliverables current. Liu discusses the spin-out and why the product form factor must change with every leap in model capability in a recent Navigators interview episode.

SPC Fund IV, $575M closed to back founders at the -1-to-0 stage: South Park Commons, the "anti-accelerator" community for technologists who haven't yet committed to a specific idea, has closed its largest fund to date, more than doubling its previous $275M vehicle. General Partner Aditya Agarwal says the raise brings SPC to $2B in AUM and a 1,200-member community, with the fund covering seed and Series A rounds alongside continued support for its strongest portfolio companies.

Standard Machines launches RL environments for chip design, aiming to let small teams tape out advanced chips in months: Chip design still takes years and hundreds of experts, while AI models improve on a cadence of months. Standard Machines, founded by ex-Apple GPU architect Jacob Peake, builds reinforcement-learning environments where models act in a real world, calling tools, editing files, and running simulations, with chip performance metrics (power, performance, area) serving as a free, automatic verifier. The goal is to compress that multi-year design cycle so small teams can take a chip from architecture to tape-out in months.